GenRecon

GenRecon is a multi-view 3D scene reconstruction approach that bridges generative 3D priors (e.g., Trellis.2) with traditional reconstruction. It divides a scene into overlapping chunks and performs conditional 3D generation for each chunk, using projected image features aligned to the prior’s 3D space, then fuses them into a complete scene. The output is an editable PBR (Physically Based Rendering) mesh, ready for lighting changes in engines like Unity or Unreal.

Compared to classic SfM/MVS or NeRF optimizations, GenRecon leverages the strong prior of generative models to fill occlusions and enhance details. It reportedly achieves 16% higher fidelity and completeness over previous state-of-the-art on indoor scenes. With just a smartphone video or a few photos, users can obtain high-quality 3D models, drastically simplifying content creation.

The project’s code and pre-trained weights have been open-sourced, but inference typically requires a capable GPU (e.g., NVIDIA RTX) for acceptable speed.